Example: Detecting Objects in Images

Overview

This is a convolutional neural network image classification example. Here, we are using a small subset of the Core 50 image dataset. The data configuration file describes the full list of images to be loaded and their class labels. The images are not included as a part of this repository; there are two options:

  • Download the 128x128 dataset from Core 50
  • If you are on the OU Supercomputer, then the images can be accessed from: ~fagg/datasets/core50

Data

  • This example contains only a single data set that is used for training (i.e., no validation or testing data sets)
  • Configuration: images.csv

Network

  • Takes in a 128x128 image with RGB color channels and classifies the image as being in one of two classes
  • The network architecture is a simple illustration of what can be implemented:
    • 4 convolutional layers, with every other one followed with a max pooling step
    • A Global Max Pool step (implicit) reduces the feature arrays down to indivdual scalars
    • 2 hidden layers combine the features together in arbitrary ways
    • THe output is 2 units (one for each class)
    • Network description: network.txt

Experiment

  • Loss: sparse categorical cross-entropy. This matches the fact that Class in the data file is an integer representation of the true class label.
  • Merics: sparse categorical accuracy
  • Configuration: experiment.txt

Experiment Suggestion

Try training with multiple objects and see if the model stays accurate or starts confusing the objects with one another.


This site uses Just the Docs, a documentation theme for Jekyll.